Snow Avalanche Frequency Estimation (SAFE): 32 years of remote hazard 1 monitoring in Afghanistan

. Snow avalanches are the predominant hazards in winter in high elevation mountains. They cause damage 10 to both humans and assets but cannot be accurately predicted. Until now, only local maps to estimate snow 11 avalanche risk have been produced. Here we show how remote sensing can accurately inventory large avalanches 12 every year at a basin scale using a 32-yr snow index derived from Landsat satellite archives. This Snow Avalanche 13 Frequency Estimation (SAFE) built in an open-access Google Engine script maps snow hazard frequency and 14 targets vulnerable areas in remote regions of Afghanistan, one of the most data-limited areas worldwide. SAFE 15 correctly detected of the actual avalanches identified on Google Earth and in the field (Probability of Detection 16 0.77 and Positive Predictive Value 0.96). A total of 810,000 large avalanches occurred since 1990 within an area 17 of 28,500 km 2 with a mean frequency of 0.88 avalanches/km²yr -1 , damaging villages and blocking roads and 18 streams. Snow avalanche frequency did not significantly change with time, but a northeast shift of these hazards 19 was evident. SAFE is the first robust model that can be used worldwide and is capable of filling data voids on 20 snow avalanche impacts in inaccessible regions.


Introduction 23
Snow avalanches are among the fastest, up to 61 m/s -1 , and therefore most dangerous natural hazards in mountain 24 areas (Louge et al., 2012). Casualties associated with avalanches are numerous; in 2021 alone, 37 fatalities occurred 25 in the US (Colorado Avalanche Information Center, 2021) and 127 in Europe (European Avalanches Warning 26 Services, 2021), but avalanche monitoring is not consistent across the globe. Most remote mountain regions and 27 communities are not systematically monitored for avalanche occurrence. Avalanche surveys amongst remote 28 villages are sparse because regions are uninhabited; however, avalanches can block connecting roads every year 29 since avalanche volumes range from hundreds to several tens of thousand cubic meters (Gubler, 1987). Where 30 weather stations exist, avalanches can be predicted based on snow depth and other weather parameters (Greene et 31 al., 2016). However, the global weather monitoring of mountainous areas is scattered and very sparse in developing 32 nations. Detecting the avalanches is a challenge and requires temporal as well as spatial data, especially for large areas. 50 Remote sensing technology, both air and spaceborne, can cover large areas at different times of the year. Indeed, 51 the frequent collection of satellite images over the same area enables the detection of changes in snow cover as 52 well as other hazards, such as floods and landslides. Until recently, the use of remote sensing in avalanche detection 53 was sparse due to low resolution, and the automation of such processes was even more difficult because of the 54 lack of relevant algorithms that can compute big data . Other remote sensing approaches 55 for avalanche detection have used radar, Lidar, and optical data. Radar satellites, such as Sentinel-1A and B, are 56 now commonly used for detecting mass movements by assessing backscatter signal changes between two time 57 periods (before and after movement) by a co-registration of the two images. Backscatter values provide 58 information on terrain roughness and any change indicates that a mass movement or a significant erosion event 59 occurred in a given area. This technology seems very promising for avalanche detection (Eckerstorfer et al., 2017;60 Malnes et al., 2015;Martinez-Vazquez and Fortuny-Guasch, 2008;Schaffhauser et al., 2008;Tompkin and Leinss, 61 2021;Yang et al., 2020). However, the acquisition of frequent radar images is too recent to use this technique to 62 detect historical avalanches. Lidar is being used in the same regard with a higher level of precision. Lidar sensors 63 measure snow depth before and after events at submeter resolutions (Prokop, 2008;Deems et al., 2013;Prokop et 64 al., 2013;Hammond et al., 2018). However, this technology remains very expensive, and the spatial coverage is 65 limited. Therefore, Lidar data are not suitable for avalanche detection at a basin scale.
Optical data are the most available data in terms of spatial and temporal resolution as well as historical data 67 archives. Thus, we used optical data to detect avalanches on a long-term basis. Landsat-5, 7 and 8 products were 68 used as their resolution (30 m, 900 m²) is sufficient to detect small avalanches . Most of 69 these data are available at a global scale. Optical sensors can detect areas covered or not covered by snow and this 70 approach has been used in multiple studies during the past decade. Manual approaches or indices have been used 71 in such studies. For example, Landsat-8 Panchromatic images (15 m) in combination with radar images were used 72 to detect avalanches in Norway (Eckerstorfer et al., 2014). Such combinations were also recently used in west 73 Greenland to map a large number of avalanches after an unprecedented snow event (Abermann et al., 2019). To 74 our knowledge, only one recent study automated the detection of avalanches using remote sensing products and 75 an open-access scripting approach (Smith et al., 2020). This study downloaded avalanches annually for a given 76 region of interest using available Landast-8 images and computed NDSI for each image. NDSI differentiated so 77 called 'supraglacial debris' from snow cover, for the date of interest. However, this approach only covers high 78 elevations areas while our study aims to detect avalanches proximate to local communities at lower elevations 79 (typically valleys). Manual and visual approaches, despite the time consuming process, can also be applied to 80 detect avalanches using high resolution images (e.g., SPOT-6), mid-resolution (e.g., Sentinel-2A and B images), 81 or even Google Earth images (Singh et al., 2020;Yariyan et al., 2020;Hafner et al., 2021). Terrain parameters like 82 slope gradient and curvature have also been added to the avalanche detection process using Digital Elevation 83 Models (DEM) combined with Landsat-8 images (Bühler et al., 2018;Singh et al., 2019). Integrated criteria are 84 therefore recommended to detect the avalanches. To our knowledge, no such studies using remote sensing have 85 been conducted in the world, especially not in Afghanistan. 86 The general objective of this study is to map annual avalanche occurrence over the past 32 years using Landsat 87 images archives in Badakhshan region, Afghanistan. Such long-term monitoring is a first attempt globally and 88 enables us to map the frequency of avalanches that impact valley communities. Thus, we used optical data to detect 89 avalanches on a long-term basis and built an open-access script in Google Engine interface: Snow Avalanche 90 Frequency Estimation (SAFE). Landsat-5, 7 and 8 products were used as their resolution (30 m, i.e., minimum 91 detectible size of 900 m²) is sufficient to detect larger avalanches (Abermann et al., 2019;Eckerstorfer et al., , 92 2014Hafner et al., 2021;Singh et al., 2019Singh et al., , 2020Smith et al., 2020;Yariyan et al., 2020). Our objective is to 93 automatically map annual avalanche occurrence over the past 32 years using Landast-5, 7 and 8 image archives in 94 the Amu Panj basin of Afghanistan. Such long-term monitoring is a first attempt globally and enables us to map 95 the frequency of avalanches that impact remote mountain valley communities. These outputs are of keen interest 96 to decision makers who can use this automated process to map avalanche hazard in the future. The most vulnerable 97 areas, villages and roads, were mapped for improve future planning. In addition, this research enables the 98 The study covers the most mountainous region of Afghanistan -Badakhshan in the Amu Panj basin located in the 104 northeast portion of the country. Average elevation is 2761 m and mean slope gradient is 21%. This region spans 105 from Bamyan Province to the Hindu Kush range, up through the Wakhan corridor in the far east of Afghanistan. 106 The summit is Nowshak Peak at an elevation of 7492 m a.s.l. The western part of Amu Panj basin is rather flat 107 and not prone to avalanches. Annual precipitation is 600 mm occurring mostly as snow between February and 108 May (Zhang et al., 2015). This terrain and precipitation characteristics lend Badakhshan very prone to avalanches. 109 The basin is large (28,580 km²), justifying automated avalanche detection to cover this area in a reasonable amount  As the aim of this study is to detect and map the annual occurrence of avalanches during the past 32 years within 148 the study area, the monitoring approach must be reasonable and transferable from year to year. Based on frequent 149 field observations and literature  , the authors noticed that avalanches can be detected 150 using the contrast between snow cover and bare cover, but the timing is perhaps the most important consideration. 151 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009  Indeed, the script is based on the assumption that snow packages exist in lowlands, especially along rivers and 152 streams, as late as May through mid-July. At this time of the year, the terrestrial snow cover has largely melted 153 and only snow packages triggered by avalanches remain. The location of those snow packages is also very critical 154 (i.e., along riverbanks). The avalanches are indeed detectable by delineating their depositional zones (not their 155 release or transition zones); in most cases these were located on river or stream banks as observed in the field 156 because the hillslopes always route snow avalanches in this direction. We cannot differentiate between dry, wet, 157 or powder snow because the process detects the remaining snow packages as avalanches in the late season (spring 158 and summer), not in winter, nor can we delineate multiple avalanches in the same depositional feature, only the 159 combined avalanches. In winter, we were not able to differentiate contrasts between snow cover and avalanches, 160 thus our focus on the late season. 161

Google Engine Interface and Code Availability 162
The concept of detecting the 'remaining snow packages in the late season' was written in Java Script using the 163

Google
Engine platform. The script SAFE is available at: 164 https://code.earthengine.google.com/?scriptPath=users%2Farnaudcaiserman%2Fexport%3ASAFE. We selected 165 Google Engine for its relatively simplicity of use and open access code, which can be used by all stakeholders 166 involved in hazard and vulnerability assessments. Additionally, internet connections in remote areas, such as 167 within the Amu Panj basin, are limited and powerful computers required to run scripts and process big data are 168 sparse. Our script can be run by anyone in a reasonable amount of time, even with a low internet capacity. As an 169 example, yearly avalanches in our study area were downloaded and mapped from Badakhshan (SAFE was 170 processed from Khorog, University of Central Asia campus, in Tajikistan  The 200 m riparian buffer was used as a mask to clip the Landsat images. Because our area of coverage 187 encompasses very different elevations, the date of snow melt is not uniform throughout the basin. Therefore, 188 distinguishing between the depositional zone and bare land requires different times depending on elevation. To 189 accomplish this, we calculated the average elevation of the snowline for the last 20 years using MODIS products. 190 To distinguish the different melt timing between highlands from lower areas, we selected the summer snowline 191 (June-July-August; JJA). The average elevation of the JJA snowline was 4420 m during the past 20 years. Two 192 masks were therefore produced: one with a river buffer in lowlands and another for highlands. Those masks are 193 only relevant if the user carefully selects the date of interest. For lowlands (below 4420 m), our time window was 194 15 May to 15 June, indicating that the script downloads and compiles all available Landsat images acquired in this 195 range and detects avalanches efficiently because during that period the terrestrial snow cover has already melted 196 and the depositional zones are easily recognised. For higher elevations (above 4420 m), snow cover melted later; 197 dates to accurately distinguish the remaining snow packages ranged from 15 June to 15 July. After many tests, it 198 was confirmed that these date ranges reproduced the desired snow conditions during the entire 32-y period. In the 199 script, users can modify these dates (line 23 and 114) to conform to local conditions. 200 201 https://doi.org/10.5194/tc-2022-15 Preprint. Discussion started: 7 February 2022 c Author(s) 2022. CC BY 4.0 License.

Snow Index Reclassification 202
After the construction of the mask, SAFE proceeds as outlined in Figure 3. NDSI is selected to detect snow 203 avalanches in the script for its transferability from one Landsat generation to another. NDSI computes a ratio 204 between VIS and SWIR bands of Landsat satellites with negative NDSI representing non-snow cover and positive 205 values indicating snow coverage (Equation 1). Three cover types were distinguished to detect avalanches at the 206 correct time: (1) bare lands; (2) water bodies; and (3)  designates snow avalanches is selected in the script. From the selected reclassification, the script removes the 217 standalone pixels because their classification might not be precise or representative of actual cover. Next, the 218 selected 'avalanche pixels' are verified into the script avoiding manual vectorization after the downloading 219 process. The vectorization procedure of avalanches is justified by the analysis steps and post-processing after 220 downloading data. Avalanche statistics, elevations, and size are extracted from vector files.  Once the data are downloaded and imported into the GIS environment, statistical analysis commences. Every year, 228 the number and areas of avalanches are calculated to quantify the evolution of avalanches. Moreover, the size of 229 the avalanches is classified. Although a generic size classification exists (Greene et al., 2016), we decided to 230 classify avalanches by size based on local conditions. We segregated four discrete categories of avalanches: small 231 (< 1000 m²); medium (1000-5000 m 2 ), large (5000-15,000 m 2 ), and very large (15,000-100,000 m 2 ). Such a 232 classification enabled us to assess the intensity of those hazards in specific locations. SAFE is not able to detect 233 the avalanches at their time occurrence, and since these hazards are detected weeks after their occurrence, their 234 size is underestimated by SAFE due to melting. However, the estimated sizes in SAFE are still useful for 235 classifying avalanches by size since large snow deposits melt slower than small snow deposits. The small 236 avalanches that occurred in winter will appear as small hazards at the time of detection and the large events as 237 large hazards since visible snow deposits can be seen on late spring. 238 239

Validation 240
The performance of SAFE in correctly detecting snow avalanche depositional zones required careful assessment. The results suggest a good reliability of SAFE (Table 2). The overall POD is 0.77 which means that SAFE 259 identified a significant number of the avalanches that impacted valley bottoms. Moreover, it seems that SAFE 260 performs better in detecting true positive avalanches, avalanches that occurred on the ground, as shown by the 261 high PPV scores (average: 0.96). SAFE almost never detected snow avalanches that did not exist. However, SAFE 262 might miss some snow avalanches due to cloud cover on the Landsat images, especially in 2001 (Table 2; (Figure 4 and 5a). SAFE inventories snow avalanches that occurred within a year and therefore 277 identifies the most vulnerable areas, but it does not aim to forecast future avalanches. During this period, some 278 810,000 snow avalanches impacted valleys within the Amu Panj basin (28,500 km²), i.e., approximately 28 279 avalanches km -2 . Each year these avalanches deposits cover an average of 1.23% of the basin area but sizes vary. 280 Avalanche size ranged from 700 to 100,000 m 2 and was categorized into four classes: small (< 1000 m²); medium 281 (1000-5000 m 2 ), large (5000-15,000 m 2 ), and very large (15,000-100,000 m 2 ). The most frequent are medium-size 282 avalanches; 342,000 events during the past 32 years. Our approach also identifies very large snow avalanches that 283 pose the greatest danger to local populations and infrastructure. We found no correlation between altitude of 284 depositional zones and avalanche size. Avalanches deposits in this region have an average altitude of 3820 m and 285 the lowest avalanche occurred at 1755 m. 286 These spatial and temporal statistics allow for a geographic assessment of the avalanches. In total, ten sub-287 catchments (ranging from 18 to 240 km 2 ) were impacted by more than one avalanche km -2 y -1 , with an average 288 frequency of 0.26 avalanches km -2 y -1 throughout the Panj Amu basin ( Figure 6). More importantly, these maps 289 prioritize villages prone to avalanches and inform relevant stakeholders which villages and infrastructure are most Our remote sensing approach facilitates innovation in snow avalanche monitoring: i.e., detecting avalanches 298 outside of populated areas, especially along roads that are frequently blocked by avalanches (Figure 8). More than 299 2000 roads in the basin (5.47% of the road network) were affected by avalanches every year. Additionally, more 300 than 400 roads in Upper Badakhshan and Wakhan regions experienced more than 2 avalanches y -1 km -1 of road 301 (within a 1 km buffer). The average frequency along roads is 0.86 avalanches km -1 y -1 during the past 32 years, 302 most of these in the medium-size category.

Stream blocking and resultant flooding 325
Damages to infrastructure and blocking of roads by avalanche depositional are not the only consequences of these mountain hazards. Because depositional zones typically reach rivers in this steep terrain, the sudden and rapid arrival of several tons of snow can temporarily block rivers inducing short-term localized flooding. By cross-checking the map of the rivers in the Amu Panj basin with SAFE outputs, it appears that 26.2% of the river network is impacted by avalanche depositional, mainly in the high mountains. During the past 32 y, 12% of the streams have been blocked by at least 10 avalanches km -1 representing a 330 significant risk for villages in floodplains. The accumulated snow mass impounds river water until it can break through releasing a large discharge surge. Thus, depending on the size of the avalanche with respect to the channel dimensions, damages to villages and farmlands may occur both upstream due to impounded water (hours to weeks) and to downstream following the sudden release of water.

Snow avalanche trends during the past 32 years
This long-term monitoring of snow avalanches facilitates the assessment of the evolution of these rapid mass movements.
During the 32 years of avalanche assessment, no significant temporal trends in impacted areas were detected (Figure 9). In addition, there was no significant trend of snow avalanche sizes (p-value > 0.05). Nevertheless, some years posed much greater risk than others. In the last 32 years, ten years have been more at risk with above-average avalanche coverage: 1990coverage: , 1991coverage: , 340 1992coverage: , 1993coverage: , 1994coverage: , 1995coverage: , 1996coverage: , 2003coverage: , 2005coverage: , 2007coverage: and 2012coverage: . In particular, 2003 had many avalanches that occupied almost 6% of the surface area of the entire basin. That year was locally noted as having heavy snowfall and farmers benefited from more snowmelt in the spring, leading to higher than average crop yields in 2003 (FAO, 2003;Guimbert, 2004). Notably, the higher risk years were also characterised by lower altitudes for avalanches. There is a slight negative correlation Pearson test) between altitude and total annual avalanche area. With larger avalanche areas, depositional reaches closer to villages. For 345 example, in 2003, the lowest avalanche occurred at an altitude of only 1871 m, very close to housing clusters and roads. It is therefore possible that communities below 2000 m are also impacted by snow avalanches and in many mountain regions of the world this represents a significant proportion of the communities living around these altitudes.

Temporal geographic shifts of snow avalanches
Long-term monitoring also shows the evolution of the spatial distribution of snow avalanches. The pattern of snow avalanches 355 has changed with time and slightly shifted to the northeast portion of the basin; thus, more avalanches are now occurring in the northeast than in the southwest ( Figure 10). Nevertheless, snow coverage did not shift simultaneously according to our remote sensing analysis nor did the snowline evolve and remained variable over the last 32 years. The geographic shift of avalanches is therefore likely due snow depth evolution. Deeper snowpacks trigger snow avalanches. There are no available data on snow depth at such a scale. However, remotely sensed land surface temperature changed during the last 20 years, with 360 a warmer band through the central portion of the basin in December (p-value 0.03 with an increase of 0.88 C°y -1 ). This central portion is mainly mountainous and this temperature pattern may have offset the avalanches to the northern mountains of the area, while the south is characterized by lower mountains. Overall, avalanche locations tend to follow the spatial distribution of snow depth . This means that despite the high variability of the snow line and snow coverage, snow avalanche distribution can significantly change over time with temperature changes and local communities must be prepared 365 for shifting hazards.

Sensitivity analysis of SAFE
To better understand how SAFE works and assess its performance, a sensitivity analysis was conducted between the model 375 parameters. The number and size of avalanches vary according to the buffer used, the dates of the Landsat images, and finally https://doi.org/10.5194/tc-2022-15 Preprint. Discussion started: 7 February 2022 c Author(s) 2022. CC BY 4.0 License. the NDSI range during the snow classification. The sensitivity analysis was conducted for the year 2019, when SAFE was most robust in valleys where actual avalanches were quite visible on Google Earth images (POD: 0.84 and PPV: 0.94). First, we run SAFE with different buffer widths (25 m of difference between each buffer). There is a strong positive correlation (0.98) between the number of avalanches detected by SAFE and the buffer width ( Figure 11A). The wider the buffer around 380 the rivers, the more avalanches SAFE will find. On the other hand, the wider the buffer, the smaller the average avalanche size (positive correlation of 0.71). This is because a large buffer extends upslope where small snow patches reside, which are not avalanches since they are located at the top of hillslopes. This means that the user should not select a buffer that is too wide, rather the area should only include the riparian zone of rivers and streams where the snow avalanche deposits are located. As such, we recommend a value of 200 m for the entire region studied. 385 The number and size of avalanches detected by SAFE depends on the NDSI range when classifying the snow. NDSI is used to differentiate between water bodies, bare lands, and snow. By varying the NDSI ranges of snow in the script, we notice a strong positive correlation with the number of avalanches detected by SAFE. The closer the index is to 0, the more hazards SAFE finds. However, this correlation shows us that the choice of NDSI range is important because we notice a threshold at 395 0.31 ( Figure 11B). Avalanches seem to be more numerous with an NDSI lower than 0.31 because the snow pixels are confused with water bodies. It is therefore essential for the user to select an NDSI higher than 0.31 to distinguish water bodies (rivers, flood areas or lakes) and snow. However, there is no correlation between the NDSI ranges and the average avalanche size as NDSI cannot interpret pixels other than 'snow' above the 0.31 threshold. Finally, the date of interest is a key parameter in SAFE. In figure 11 C, we can see that the number of avalanches detected by SAFE is highest at the end of winter due to the 400 almost constant cloud cover since January, but also due to the inability to distinguish avalanches from snow cover in winter (with Landsat images). May is a key month in SAFE applications: the snow coverage, which is thinner than the avalanche packages, begins to melt and the number of avalanches detected can then be assessed. It is therefore essential to select post-May images to detect avalanches, while taking care not to select post-July images as avalanches melt in summer and make their detection impossible. 405 Some avalanches deposits can also be visible on successive images, after the snow coverage melted. SAFE was specifically designed to detect avalanches at their earliest stage after snow coverage melted. Indeed, starting from May (when the avalanches are not confused anymore with snow coverage), the snow deposits will start to melt and the sizes to be underestimated. For that reason, it is important to select late spring images for lowlands avalanches and early summer for high lands avalanches, not later. Cloud cover is another issue in avalanches locations and sizes detection since they can partly or 410 fully cover the avalanches at the time of the image. This is another reason to select images starting from late spring when the cloud coverage of this region is at its lowest and even null in early summer. If the cloud coverage is high even late spring, the users can still select later images, but the risk is to detect avalanches deposits that have significantly started to melt.
To summarize, we recommend the following three parameters in the SAFE script: buffer of 200 m to include only snow avalanche deposits; NDSI > 0.31 to distinguish water bodies from snow, and images from May to July to distinguish avalanches 415 from snow cover.

Excluding snow coverage
Interpreting the remaining snow packages as depositional avalanches can lead to some errors. Indeed, despite a precise masking operation (excluding summits and very high plateaus where snow persists), in some cases the use of NDSI might not properly 420 segregate avalanches from large areas of remaining snow. After assessing the sizes of true avalanches (the ones that SAFE correctly detected based on Google Earth images), it appeared that snow covers > 100,000 m² were not avalanches but snow coverage and thus we removed them. However, in highlands, even along riverbanks, some snow packages interpreted as avalanches may be remaining snow cover. As such, the date range was selected as late as possible in the year. Thus, it is https://doi.org/10.5194/tc-2022-15 Preprint. Discussion started: 7 February 2022 c Author(s) 2022. CC BY 4.0 License. advised to keep the mask at the very bottom of valleys (maximum 500 m buffer along the river) to exclude high plateaus and 425 potential snow-covered areas.

Water bodies in SAFE
A final limitation of using this remote sensing and NDSI approach for avalanche detection is the possible confusion between some small water bodies and avalanches. Indeed, in some cases certain river reaches (order > 4 in our study) could be 430 interpreted as snow because they were frozen and appeared as white pixels on the Landsat archives. The same issue can occur with ponds and lakes. This limitation was foreseen before processing the images in our study and we excluded these large water bodies from the region of interest (in the mask) by using available shapefiles. For example, Shiva Lake, one of the largest water bodies in Amu Panj basin (15 km²), was removed from the analysis. Another way to avoid the water pixel selection is to adapt the NDSI reclassification itself, depending on the study area. This is possible lines 51-54 for low elevations and lines 435 142-144 for high elevations in the script.

SAFE outcomes compared to other snow avalanches detection studies
SAFE contributes to the literature on snow avalanche detection, but in a unique way using remote sensing. As noted, many studies and models exist using various products: Radar, Optical, and Topographic. The strength of remote sensing is the 440 automatic processing at a large scale and over long timeframes. SAFE uses the capabilities of remote sensing by processing more than one image per year at the catchment scale. Moreover, the use of Landsat archives allows assessment over the last 32 years, which is not yet possible with recent Radar data such as Sentinel-1. Most of the current avalanche detection models use freely available products, with acceptable if not good accuracy (Table 3). The accuracy of these studies using Radar images ranges from 53 to 81% making this a relatively robust tool. One of the reasons why SAFE does not use Radar images is the 445 weight of the images (data storage), especially Sentinel-1, which is mostly above 1 Gb/image. The heaviness of these images is not suitable for a model like SAFE, which was specifically designed for remote study areas where internet connections may be very limited. Other models also exist with Optical images with high accuracy ranging from 71 to 93% (Table 3). In the optical domain, SAFE showed a POD of 77% over an area of 28,500 km². SAFE is therefore in the high range of models with optical, medium resolution (Landsat) images. Furthermore, one of the perspectives of this paper would be to compare the 450 results of the different models in Table 3 with the results of SAFE over the same area where data on the size and location of avalanches would be available.

Conclusion
SAFE can be considered as a universal approach to assess snow avalanches where ground data are very limited, such as in the Afghan mountains. Here we showed the capability of long-term remote sensing data to robustly detect snow avalanches that impact valley locations. While we have successively applied SAFE to assess avalanche frequency and impacts in valleys and 465 lower hillslopes of Afghanistan, arguably one of the most data-limited regions worldwide, this model should perform even better in areas where snow data are available making it an important tool for avalanche vulnerability assessment worldwide.
More than 30 years after the launch of Landsat-5, it is now possible to compile all data and assess the temporal as well as spatial evolution of such hazards. NDSI is a relevant index to detect avalanches provided the correct region and dates of interest are selected -i.e., riverbanks during the late melt season. The thickness of the depositional zones facilitates the detection of 470 avalanches after the snow cover has melted on hillslopes in spring or early summer.
The automation of snow avalanche detection using remote sensing technologies at regional scales is still new and SAFE was designed to guide decision-makers, planners, and disaster risk practitioners. Indeed, such people can now know where the most at-risk areas are located based on these frequency maps. Such information informs the relative risk of building sites and land use decisions in such mountainous terrain with greater precision. The level of exposure of roads to avalanches can also be 475 estimated using these frequency maps, and can inform road managers regarding road location, maintenance practices, and mitigation structures. The tourism sector can also benefit from this snow avalanche inventory, especially the winter sports industry. Furthermore, this method can also be used to prioritize areas for more sophisticated and data-intensive avalanche risk analysis (Keylock et al., 1999). SAFE can be applied by any user throughout mountainous regions of the world as it is designed to be user-friendly. Thus, frequent users can contribute to the robustness of the snow avalanche archive, thus improving 480 recommendations to policy makers. https://doi.org/10.5194/tc-2022-15 Preprint. Discussion started: 7 February 2022 c Author(s) 2022. CC BY 4.0 License.